一种基于树型贝叶斯网络的集成多标记分类算法

在多标记分类问题中,有效地利用标记间的依赖关系是进一步提升分类器性能的主要途径之一。基于分类器链算法,利用互信息度量理论构造分类对象的类属性之间明确的多标记关系依赖模型,并依据建立的标记依赖模型将分类器链中的线性依赖拓展成树型依赖,以适应更为复杂的标记依赖关系；同时,在此基础上利用Stacking集成学习方法建立最终训练模型,提出了一种新的针对树型依赖表示模型的Stacking算法。 在多个实验数据集上的实验结果表明,与原有的Stacking集成学习相比,该算法提升了分类器的相应评价指标。

Ensemble Multi-label Classification Algorithm Based on Tree-Bayesian Network

The performance of learning algorithm can be improved by utilizing existing label dependencies in multi-label classification.Based on the strategy of classifier chain and stacking ensemble learning,this paper built a model to explain the dependency of different labels,and extended the linear dependency into tree dependency to deal with much more complicated label relations.Compared with the original Stacking algorithm,the performance of the proposed algorithm is improved in the experiments.